Detection Accuracy Performance Research Articles

Early detection of chronic kidney disease using eurygasters optimization algorithm with ensemble deep learning approach

Chronic kidney disease (CKD) emerges as a global health problem with high morbidity and mortality rates, and it induces other diseases. Patients often fail to notice these diseases because there are no obvious symptoms during the earlier stage of CKD. Earlier diagnosis of CKD allows them to receive early medical intervention to improve the disease progression and outcome. Earlier stratification of CKD can affect the health care provided to patients through various options, such as kidney transportation, hemodialysis, or pharmacological care in milder cases. Recently, deep learning (DL) and machine learning (ML) approaches have gained significance in the domain of medical diagnoses due to their high prediction accuracy. DL and ML approaches can successfully aid physicians in achieving these goals due to their accurate and fast detection performance. The performance of the proposed technique mainly depends on selecting the suitable algorithms and appropriate features. Thus, the study introduces an Eurygasters Optimization Algorithm with an Ensemble DL CKD detection (EOAEDL-CKDD) method. The EOAEDL-CKDD method aims to detect and classify the presence of CKD using feature selection and hyperparameter tuning strategies. The EOAEDL-CKDD method applies a min-max scalar to convert the data into a uniform format to achieve this. The EOAEDL-CKDD technique exploits EOA to select features. This is followed by an ensemble of long short-term memory (LSTM), bidirectional gated recurrent unit (BiGRU), and bidirectional LSTM (BiLSTM) models that are used for the CKD detection process. Finally, a shuffled frog leap algorithm (SFLA) based hyperparameter selection process is carried out to choose the ensemble models' hyperparameter values optimally. The empirical assessment of the EOAEDL-CKDD method is tested on the benchmark CKD dataset. The experimental values highlighted that the EOAEDL-CKDD technique gains an optimal detection rate compared to existing models.

Online insulator defects detection and application based on YOLOv7-tiny algorithm

As an indispensable part of the power transmission system, insulators are of great importance to the safe and stable operation of power grids in terms of their healthy and reliable operation. To realize real-time monitoring of insulator defects under a complex environment, this paper proposes an insulator defect detection method based on the You Only Look Once version 7-tiny (YOLOv7-tiny) algorithm. Then an edge device-unmanned aerial vehicle (UAV) inspection system is developed to verify the real-time performance of the algorithm. By introducing the structure intersection over union (SIoU) loss function to the YOLOv7-tiny model, the regression speed of the anchor frame can be effectively accelerated on the basis of the miniature model, to accelerate the model operation. Thereafter, a smooth sigmoid linear unit (SiLU) activation function is used in the network neck to improve the nonlinear representation ability; After that, an edge computing device based on NVIDIA Jetson Xavier NX is established to verify the real-time performance of the method. Experimental results reveal mean average precision (mAP) of insulators and their missing series defects is as high as 98.31%. Besides, the detection speed of the designed model deployed to mobile edge devices can reach 35 frames per second (FPS), with real-time and accurate detection performance of insulators and their missing series defects.

Achievement of dynamic tablet defect detection mechanism using biaxial slope symmetry algorithm

An inspection in tablet appearance integrity before bottling is regarded as a routine task in a pharmaceutical factory. Although some methods such as automated optical instrument, video or artificial intelligence (AI) are currently available in industry, it usually pays for a complex computational process as well as high cost. Based on the symmetry of tablet appearance in reality, this study develops a biaxial scanning slope symmetry algorithm to realize a dynamic real-time tablet defect detection with a simple arithmetic operation. First, the tablet is discretely scanned using image sensor in two axes, i.e. X and Y directions, simultaneously. Second, the analogy output signals generated from the sensor during the scanning process is discretely digitized and stored in an array. Third, the coordinate of center point in the series data array is identified from every line scanning. Fourth, every section slope between two nearby center points from the first to last lines is formulated and calculated sequentially. Finally, the square mean error (SME) is used to evaluate the shape defect situation according to all accumulated errors from every slope variation. The experimental results verify that the proposed algorithm can achieve both fast and accurate detection performance.

ResMon: Domain-Adaptive Wireless Respiration State Monitoring via Few-Shot Bayesian Deep Learning

Under the outbreak of the COVID-19 pandemic, respiration state monitoring plays an important role in assisting respiratory disease diagnosis and treatment. Thanks to the non-intrusive nature and low deployment cost, Wi-Fi based wireless respiration state monitoring methods have gained increasing popularity. By analyzing the variation of channel state information (CSI) of Wi-Fi signals, the respiration states of a target person under the wireless coverage, such as cough, sneeze, and yawn, can be accurately detected. A major problem of the current wireless respiration state monitoring methods is being overly domain-dependent. That is, a sensing algorithm fine-tuned to a specific device placement and background setting (i.e., a domain) can result in drastic drop in detection accuracy when applied to a dissimilar new domain. To enhance the robustness of wireless sensing and reduce the sensing cost across different domains, we propose in this paper a domain-adaptive Respiration state Monitoring system (ResMon) that achieves highly accurate cross-domain detection performance while requiring very limited labeled samples in the new domain. In a nutshell, the proposed ResMon consists of a source domain meta-training stage and a target domain meta-testing stage. In the meta-training stage, we leverage the rich source domain labeled dataset to train an embedding model as a feature extractor of high-dimensional CSI data measurements. In particular, we apply the statistical Bayesian deep learning technique to improve the generalization performance of the embedding model in cross-domain applications. In the meta-testing stage, we combine the embedding model with a few-shot learning technique to train a domain-specific classifier using very limited labeled samples in the target domain. Experiment results show that the proposed ResMon can achieve on average 87.26% cross-domain detection accuracy in a 4-class respiration state classification task using only 5 labeled samples per class, which significantly outperforms the considered benchmark methods.

Surveying wildlife and livestock in Uganda with aerial cameras: Deep Learning reduces the workload of human interpretation by over 70%

As the need to accurately monitor key-species populations grows amid increasing pressures on global biodiversity, the counting of large mammals in savannas has traditionally relied on the Systematic-Reconnaissance-Flight (SRF) technique using light aircrafts and human observers. However, this method has limitations, including non-systematic human errors. In recent years, the Oblique-Camera-Count (OCC) approach developed in East Africa has utilized cameras to capture high-resolution imagery replicating aircraft observers’ oblique view. Whilst demonstrating that human observers have missed many animals, OCC relies on labor-intensive human interpretation of thousands of images. This study explores the potential of Deep Learning (DL) to reduce the interpretation workload associated with OCC surveys. Using oblique aerial imagery of 2.1 hectares footprint collected during an SRF-OCC survey of Queen Elizabeth Protected Area in Uganda, a DL model (HerdNet) was trained and evaluated to detect and count 12 wildlife and livestock mammal species. The model’s performance was assessed both at the animal instance-based and image-based levels, achieving accurate detection performance (F1 score of 85%) in positive images (i.e. containing animals) and reducing manual interpretation workload by 74% on a realistic dataset showing less than 10% of positive images. However, it struggled to differentiate visually related species and overestimated animal counts due to false positives generated by landscape items resembling animals. These challenges may be addressed through improved training and verification processes. The results highlight DL’s potential to semi-automate processing of aerial survey wildlife imagery, reducing manual interpretation burden. By incorporating DL models into existing counting standards, future surveys may increase sampling efforts, improve accuracy, and enhance aerial survey safety.

Preparation of a recyclable composite aerogel with controllable peroxidase-like activity for on-site and visual analysis of the dihydroxybenzene isomers

Colorimetric detection methods have been attracted and preferred for analytes assessment. Although simple and highly efficient colorimetric strategies based on nanozymes have been developed, there remain inherent problems of intrinsic color interference for accurate analysis and erratic activity to lose controllable, recyclable, and sensitive efficacies. Herein, we have prepared a novel composite aerogel named PACO that is rationally assembled by the poly (vinyl alcohol) aerogel (PVA aerogel) and copper oxide nanoparticles (CuO NPs) to address these challenges. PACO has been proved to exhibit the controllable peroxidase (POD)-like activity through separating the solid PACO from the aqueous solutions. Meanwhile, PACO can avoid intrinsic color interferences and provide more accurate detection performance for naked-eye discrimination. Furthermore, PACO shows the recyclable and long lifetime performances for sensitive detection of water pollutants including hydroquinone (HQ) and catechol (CC) with the LOD of 0.084 μM and 0.101 μM, respectively. Finally, we have constructed a smartphone-based colorimetric detection platform to realize the on-site and visualanalysis of water pollutants in real water samples. Overall, PACO achieves the advanced detection performances to exhibit great potential in nanozymes based sensing method development and colorimetric-related analytical research.

Ensemble Transfer Learning Based Cross-Domain UAV Actuator Fault Detection

The actuator is critical for controlling the unmanned aerial vehicle (UAV) attitude. With the development of sensing technology, the convenience of collecting high-quality data by sensors has provided favorable conditions for data-driven based UAV actuator fault detection methods. However, there are less historical data on a new type of UAV, making it challenging to accurately obtain data-driven fault detection models. The data domain distribution of the new UAV (target domain) is different from that of existing UAVs (source domains), and the basic data-driven methods are difficult to achieve cross-domain fault detection. The transfer learning of a single source domain method can solve this problem, but it has a strong dependence on the single source domain data characteristics. Therefore, we propose the ensemble transfer learning method, which learns multiple UAV data characteristics to address the accuracy degradation issue caused by single-source domain insufficient characteristics and realize cross-domain actuator fault detection with limited data for the new type of UAV. In the method, the basic actuator prediction models are established based on the bidirectional long short-term memory (BiLSTM) algorithm. Then we combine the transfer learning method to realize the cross-domain fault detection model establishment of the new type of UAV. In addition, a cross-domain similarity metric based on R-score is creatively proposed for the ensemble to avoid excessive dependence on a single-source domain. We conduct experiments based on the ArduPilot open-source simulation platform, and the results show that our proposed method can adaptively obtain more accurate fault detection performance.

An improved YOLOX approach for low-light and small object detection: PPE on tunnel construction sites

Abstract Tunnel construction sites pose a significant safety risk to workers due to the low-light conditions that can affect visibility and lead to accidents. Therefore, identifying personal protective equipment (PPE) is critical to prevent injuries and fatalities. A few researches have addressed the challenges posed by tunnel construction sites whose light conditions are lower and images are captured from a distance. In this study, we proposed an improved YOLOX approach and a new dataset for detecting low-light and small PPE. We modified the YOLOX architecture by adding ConvNeXt modules to the backbone for deep feature extraction and introducing the fourth YOLOX head for enhancing multiscale prediction. Additionally, we adopted the CLAHE algorithm for augmenting low-light images after comparing it with eight other methods. Consequently, the improved YOLOX approach achieves a mean average precision of 86.94%, which is 4.23% higher than the original model and outperforms selected state-of-the-art. It also improves the average precision of small object classes by 7.17% on average and attains a real-time processing speed of 22 FPS (Frames Per Second). Furthermore, we constructed a novel dataset with 8285 low-light instances and 6814 small ones. The improved YOLOX approach offers accurate and efficient detection performance, which can reduce safety incidents on tunnel construction sites.

Biologically Inspired Machine Learning-Based Trajectory Analysis in Intelligent Dispatching Energy Storage System

The present work expects to explore the application effect of biologically inspired Plasticity Neural Network in the industrial intelligent dispatching energy storage system, and highlight the intelligence and fault detection performance of the control system. To address the faults in intelligent dispatching energy storage system, the present work implements a fault diagnosis model of intelligent dispatching energy storage system based on Deep Belief Network (DBN), and simulates and analyzes the model. The results show that the transmission probability of the fault diagnosis model of the constructed intelligent energy storage scheduling system is 100% and when the parameters <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\lambda $ </tex-math></inline-formula> is between 0.01 and 0.05, the real-time performance of data transmission is the highest. Compared with other classical algorithm models, the success rate and detection accuracy of the proposed algorithm are about 85%, the energy consumption is lower, and the detection effect is more obvious. Therefore, the constructed system obviously has higher real-time performance and more accurate fault detection performance, and significantly better system detection and protection performance. The results provide an experimental basis for the operation and fault detection of intelligent dispatching energy storage system.

Hierarchical LSTM-Based Network Intrusion Detection System Using Hybrid Classification

Most existing network intrusion detection systems (NIDSs) perform intrusion detection using only a partial packet data of fixed size, but they suffer to increase the detection rate. In this study, in order to find the cause of a limited detection rate, accurate intrusion detection performance was analyzed by adjusting the amount of information used as features according to the size of the packet and length of the session. The results indicate that the total packet data and all packets in the session should be used for the maximum detection rate. However, existing NIDS cannot be extended to use all packet data of each session because the model could be too large owing to the excessive number of features, hampering realistic training and classification speeds. Therefore, in this paper, we present a novel approach for the classifier of NIDSs. The proposed NIDS can effectively handle the entire packet information using the hierarchical long short-term memory and achieves higher detection accuracy than existing methods. Performance evaluation confirms that detection performance can be greatly improved compared to existing NIDSs that use only partial packet information. The proposed NIDS achieves a detection rate of 95.16% and 99.70% when the existing NIDS show the highest detection rate of 93.49% and 98.31% based on the F1-score using two datasets. The proposed method can improve the limitations of existing NIDS and safeguard the network from malicious users by utilizing information on the entire packet.

Can you see me lying? Investigating the role of deception on gaze behavior

Throughout the early part of this century, and especially during the peak of the global pandemic of 2020, the world has come to rely increasingly on computer-mediated communication (CMC). The study of computer-based media and their role in mediating communication has long been a part of the academic study of information systems. Unfortunately, human communication, regardless of the medium over which it occurs, involves deception. Despite the growing reliance on CMC for communication, a limited amount of work has considered deception and its detection in mediated environments. The study reported here investigates the communication issues associated with cue restrictions in CMC, specifically videoconferencing, and with how these restrictions affect deception detection success. We employed eye tracking technology to analyze the visual behavior of veracity judges and how it influenced their assessments. We found that the visual foci of the judges varied as a result of the message veracity. Judges fixated longer on the mouth and torso of speakers when messages were deceptive and focused longer on the hands of the speakers when messages were truthful. We also found that fixating longer on the mouth and torso of the speakers was associated with less credible assessment of the speakers. Last, longer gaze fixations on the torso and left hand of the speakers resulted in less accurate deception detection performance.

A multi-source information fusion model for outlier detection

Multi-source information fusion (MSIF) is a useful strategy for combining complimentary data from numerous information sources to produce an overall precise description, which can help with effective decision-making, prediction, and categorization, etc. In order to find the objects that are different from the expected ones after fusion, i.e., anomalies, or outliers, an MSIF model is put forward for outlier detection. This is a two-stage model that includes fusion of multiple information sources and outlier detection of fused data. The first stage uses information sets to construct uncertainty criteria for information source values and combines multiple information sources into a single information source based on the minimum uncertainty strategy. The second stage uses the Gaussian kernel method for possibility modeling based on the fused data to construct knowledge granules. From the perspective of granular computing, outliers in the fused data can be assigned to each knowledge granule. Then, we can find all outliers just by evaluating these knowledge granules. Inspired by this, the fuzzy knowledge measure (FKM) is proposed to evaluate the knowledge granule. Moreover, several metrics are induced on the basis of FKM to describe outliers in knowledge granules and an FKM-based outlier detection algorithm (FKMOD) is designed. Finally, we conduct the experiments on sixteen open access outlier detection datasets. The experimental results show that the proposed FKMOD method has more accurate detection performance than nine classical methods.

A microfluidic-chip-based system with loop-mediated isothermal amplification for rapid and parallel detection of Trichomonas vaginalis and human papillomavirus.

Cervical cancer is a significant global health issue primarily caused by high-risk types of human papillomavirus (HPV). Recent studies have reported an association between Trichomonas vaginalis (T. vaginalis) infections and HPV infections, highlighting the importance of simultaneously detecting these pathogens for effective cervical cancer risk management. However, current methods for detecting both T. vaginalis and HPV are limited. In this study, we present a novel approach using a microfluidic-chip-based system with loop-mediated isothermal amplification (LAMP) for the rapid and parallel detection of T. vaginalis, HPV16, HPV18, and HPV52 in a reagent-efficient and user-friendly manner. Compared to conventional LAMP assays in tubes, our system exhibits enhanced sensitivity with values of 2.43 × 101, 3.00 × 102, 3.57 × 101, and 3.60 × 102 copies per reaction for T. vaginalis, HPV16, HPV18, and HPV52, respectively. Additionally, we validated the performance of our chip by testing 47 clinical samples, yielding results consistent with the diagnostic methods used by the hospital. Therefore, our system not only offers a promising solution for concurrent diagnosis of T. vaginalis and HPV infections, particularly in resource-limited areas, due to its cost-effectiveness, ease of use, and rapid and accurate detection performance, but can also contribute to future research on the co-infection of these two pathogens. Moreover, the system possesses the capability to simultaneously detect up to 22 different types of pathogens, making it applicable across a wide range of domains such as diagnostics, food safety, and water monitoring.

Deep learning-based detection of aluminum casting defects and their types

Due to its unique properties, high-pressure aluminum die-casting parts are used quite often, especially in the automotive industry. However, die-casting is a process which requires non-destructive testing of the critical components using technologies such as X-ray to examine the internal defects that are not otherwise visible. Such a timeconsuming visual inspection requires well-trained human specialists with the utmost attention. In this study, state-of-the-art deep learning-based object detection methods were trained using an X-ray image dataset of aluminum parts to detect internal defects and predict their types without human attention. The Al-Cast image dataset used in this study contains 3466 images of parts produced in high-pressure die casting machines. It is shared as an open-access original database for the nondestructive testing (NDT) community. ASTM standard definitions for aluminum casting defects are used in determining their types, and to the best of our knowledge, this novel approach is the first in the deep learning literature. Among the 12 deep learning-based object detection methods used for comparison, YOLOv5 versions yielded the highest detection accuracy (0.956 mAP) with the shortest training time (0.75 h). In addition, tests were performed for both original and contrast enhanced images on 348 test images. YOLOv5m performed an accurate detection performance of 95.9%. Additionally, YOLOv5n can process 132 images per second. This study can be considered the first step of an artificial intelligence product that can detect internal defects of aluminum casting parts with industrial standards and explain the relationship between highpressure injection die casting parameters and these defects.

Rapid Harmonic Detection Scheme Based on Expanded Input Observer

The existence of harmonics will cause the quality of power supply in a power system to decline and will affect the normal use of the power system. Therefore, it is important to suppress harmonics in the power system, and the first step of harmonic suppression is harmonic detection. To address this phenomenon, a fast harmonic detection method is proposed in this paper. It is based on the input observer theory to construct a state space model based on the original signal and harmonic components and estimate the state variables so as to achieve harmonic extraction. The characteristic roots are used to prove the convergence of the observer. In addition, the Second-Order Generalized Integration (SOGI) frequency estimation method is chosen to cascade with it so that harmonic detection can be accomplished with unknown frequencies. The simulation results prove that the proposed method can quickly converge and accurately extract each harmonic in the case of fluctuations in the fundamental amplitude, fundamental frequency and phase of the input signal, and the whole process can be completed in 0.02 s. The possible effects of white noise on harmonic extraction are also simulated, and the results prove that the accuracy of harmonic extraction can still be guaranteed in the presence of white noise. By using the Speedgoat real-time target machine built Rapid Control Prototype (RCP) as a testbed, experiments with similar simulation conditions were performed. The results show that the method has fast and accurate harmonic detection performance.

IR-Net: An improved RetinaNet-based ship detection detector

The task of ship object detection is of great importance to research. Ship object at sea is key to marine monitoring, real-time rescue, and wartime attack. Whether the ship object can be identified quickly and accurately to support the commander's decision-making is largely related to the success of the maritime mission. However, the dense arrangement and arbitrary direction of ships make it more difficult than the general object detection. Therefore, this paper proposes a rotation detector to deal with these problems. Furthermore, ship detection faces some chanllenges such as the complex background of offshore port images, the large ratio of ship length to width, and the potential of being blocked by clouds. Therefore, this paper proposes a rotatable bounding box based on RetinaNet, named Improved RetinaNet (IR-Net). With the deeper ResNet as its backbone network, this network extracts deeper object features. Furthermore, DCM (deformable convolutional module) is added to the network, so that the sampling points of the network can change with the shape of the object ship, thus extracting the object features more effectively. Finally, in this paper, the Cutout data enhancement strategy in the network is added, which effectively improves its detection accuracy in the case of object occlusion. Ablation experiments show that the IR-Net proposed in this paper has a more accurate detection performance on the HRSC2016 ship dataset.

Design and fabrication of an ultra-sensitive Ta2C MXene/Au-coated tilted grating sensor

Accurate evaluation of trace biochemical molecular signals by fiber-optic sensors is highly challenging. Recently, MXene, as one of the most promising 2D materials with unique optoelectronic properties and structural advantages, has attracted numerous concerns about the sensitization of optical fiber sensors. This paper proposed an innovative Ta2C MXene (Ta2C-M) sensitized Au-coated tilted fiber Bragg grating (TFBG) SPR sensor. We achieved the optimal deposition scheme for Ta2C-M by varying the laser induction time. The introduction of Ta2C-M brought a 23.3% increase in sensitivity and exhibited an excellent figure of merit (FOM=2586.11 RIU-1) and linearity (R2 =0.99903). Based on COMSOL, a 3D diversity analysis model was proposed, which validated the polarization response of the TFBG sensor, the electromagnetic gain of the Ta2C-M coating and enabled the fitting evaluation of the SPR behavior. The sensitization mechanism of the Ta2C-M coating was further analyzed utilizing energy band theory based on the Vienna ab initio Simulation Package (VASP). The assessment of two carcinogenic food additives, Saccharin and BHT, demonstrated the extreme SPR signal resolution and accurate trace detection performance (10-11 M) of the Ta2C-M/Au/TFBG sensor. These findings indicate that the proposed sensor will hopefully open new biotechnology, medicine, and food safety detection avenues.

Automatic Detection of Osteochondral Lesions of the Talus via Deep Learning

Screening of osteochondral lesions of the talus (OLTs) from MR imags usually requires time and efforts, and in most case lesions with small size are often missed in clinical practice. Thereore, it puts forward higher requirements for a more efficient OLTs screening method. To develop an automatic screening system for osteochondral lesions of the talus (OLTs), we collected 92 MRI images of patients with ankle pain from Qilu Hospital of Shandong University and proposed an AI (artificial intelligence) aided lesion screening system, which is automatic and based on deep learning method. A two-stage detection method based on the cascade R-CNN model was proposed to significantly improve the detection performance by taking advantage of multiple intersection-over-union thresholds. The backbone network was based on ResNet50, which was a state-of-art convolutional neural network model in image classification task. Multiple regression using cascaded detection heads was applied to further improve the detection precision. The mean average precision (mAP) that is adopted as major metrics in the paper and mean average recall (mAR) was selected to evaluate the performance of the model. Our proposed method has an average precision of 0.950, 0.975, and 0.550 for detecting the talus, gaps and lesions, respectively, and the mAP, mAR was 0.825, 0.930. Visualization of our network performance demonstrated the effectiveness of the model, which implied that accurate detection performance on these tasks could be further used in real clinical practice.

Sound Event Detection by Pseudo-Labeling in Weakly Labeled Dataset.

Weakly labeled sound event detection (WSED) is an important task as it can facilitate the data collection efforts before constructing a strongly labeled sound event dataset. Recent high performance in deep learning-based WSED’s exploited using a segmentation mask for detecting the target feature map. However, achieving accurate detection performance was limited in real streaming audio due to the following reasons. First, the convolutional neural networks (CNN) employed in the segmentation mask extraction process do not appropriately highlight the importance of feature as the feature is extracted without pooling operations, and, concurrently, a small size kernel forces the receptive field small, making it difficult to learn various patterns. Second, as feature maps are obtained in an end-to-end fashion, the WSED model would be weak to unknown contents in the wild. These limitations would lead to generating undesired feature maps, such as noise in the unseen environment. This paper addresses these issues by constructing a more efficient model by employing a gated linear unit (GLU) and dilated convolution to improve the problems of de-emphasizing importance and lack of receptive field. In addition, this paper proposes pseudo-label-based learning for classifying target contents and unknown contents by adding ’noise label’ and ’noise loss’ so that unknown contents can be separated as much as possible through the noise label. The experiment is performed by mixing DCASE 2018 task1 acoustic scene data and task2 sound event data. The experimental results show that the proposed SED model achieves the best F1 performance with 59.7% at 0 SNR, 64.5% at 10 SNR, and 65.9% at 20 SNR. These results represent an improvement of 17.7%, 16.9%, and 16.5%, respectively, over the baseline.

Neural network fast-classifies biological images through features selecting to power automated microscopy.

Artificial intelligence is nowadays used for cell detection and classification in optical microscopy during post-acquisition analysis. The microscopes are now fully automated and next expected to be smart by making acquisition decisions based on the images. It calls for analysing them on the fly. Biology further imposes training on a reduced data set due to cost and time to prepare the samples and have the data sets annotated by experts. We propose a real-time image processing compliant with these specifications by balancing accurate detection and execution performance. We characterised the images using a generic, high-dimensional feature extractor. We then classified the images using machine learning to understand the contribution of each feature in decision and execution time. We found that the non-linear-classifier random forests outperformed Fisher's linear discriminant. More importantly, the most discriminant and time-consuming features could be excluded without significant accuracy loss, offering a substantial gain in execution time. It suggests a feature-group redundancy likely related to the biology of the observed cells. We offer a method to select fast and discriminant features. In our assay, a 79.6 2.4% accurate classification of a cell took 68.7 3.5 ms (mean SD, 5-fold cross-validation nested in 10 bootstrap repeats), corresponding to 14 cells per second, dispatched into eight phases of the cell cycle, using 12 feature groups and operating a consumer market ARM-based embedded system. A simple neural network offered similar performances paving the way to faster training and classification, using parallel execution on a general-purpose graphic processing unit. Finally, this strategy is also usable for deep neural networks paving the way to optimizing these algorithms for smartmicroscopy.